src/runtime/thread.c

   1 /*
   2  * This software is part of the SBCL system. See the README file for
   3  * more information.
   4  *
   5  * This software is derived from the CMU CL system, which was
   6  * written at Carnegie Mellon University and released into the
   7  * public domain. The software is in the public domain and is
   8  * provided with absolutely no warranty. See the COPYING and CREDITS
   9  * files for more information.
  10  */
  11
  12 #include "sbcl.h"
  13
  14 #include <stdlib.h>
  15 #include <stdio.h>
  16 #include <string.h>
  17 #ifndef LISP_FEATURE_WIN32
  18 #include <sched.h>
  19 #endif
  20 #include <signal.h>
  21 #include <stddef.h>
  22 #include <errno.h>
  23 #include <sys/types.h>
  24 #ifndef LISP_FEATURE_WIN32
  25 #include <sys/wait.h>
  26 #endif
  27
  28 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
  29 #include <mach/mach.h>
  30 #include <mach/mach_error.h>
  31 #include <mach/mach_types.h>
  32 #endif
  33
  34 #include "runtime.h"
  35 #include "validate.h"           /* for BINDING_STACK_SIZE etc */
  36 #include "alloc.h"
  37 #include "thread.h"
  38 #include "arch.h"
  39 #include "target-arch-os.h"
  40 #include "os.h"
  41 #include "globals.h"
  42 #include "dynbind.h"
  43 #include "genesis/cons.h"
  44 #include "genesis/fdefn.h"
  45 #include "interr.h"             /* for lose() */
  46 #include "gc-internal.h"
  47
  48 #ifdef LISP_FEATURE_WIN32
  49 /*
  50  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
  51  * so define it arbitrarily
  52  */
  53 #define SIGSTKSZ 1024
  54 #endif
  55
  56 #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
  57 #define DELAY_THREAD_POST_MORTEM 5
  58 #define LOCK_CREATE_THREAD
  59 #endif
  60
  61 #ifdef LISP_FEATURE_FREEBSD
  62 #define CREATE_CLEANUP_THREAD
  63 #define LOCK_CREATE_THREAD
  64 #endif
  65
  66 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  67
  68 #ifdef LISP_FEATURE_SB_THREAD
  69 struct thread_post_mortem {
  70 #ifdef DELAY_THREAD_POST_MORTEM
  71     struct thread_post_mortem *next;
  72 #endif
  73     os_thread_t os_thread;
  74     pthread_attr_t *os_attr;
  75     os_vm_address_t os_address;
  76 };
  77
  78 #ifdef DELAY_THREAD_POST_MORTEM
  79 static int pending_thread_post_mortem_count = 0;
  80 pthread_mutex_t thread_post_mortem_lock = PTHREAD_MUTEX_INITIALIZER;
  81 #endif
  82 static struct thread_post_mortem * volatile pending_thread_post_mortem = 0;
  83 #endif
  84
  85 int dynamic_values_bytes=TLS_SIZE*sizeof(lispobj);  /* same for all threads */
  86 struct thread *all_threads;
  87 extern struct interrupt_data * global_interrupt_data;
  88
  89 #ifdef LISP_FEATURE_SB_THREAD
  90 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  91 #ifdef LOCK_CREATE_THREAD
  92 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  93 #endif
  94 #ifdef LISP_FEATURE_GCC_TLS
  95 __thread struct thread *current_thread;
  96 #endif
  97 #endif
  98
  99 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 100 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
 101 #endif
 102
 103 static void
 104 link_thread(struct thread *th)
 105 {
 106     if (all_threads) all_threads->prev=th;
 107     th->next=all_threads;
 108     th->prev=0;
 109     all_threads=th;
 110 }
 111
 112 #ifdef LISP_FEATURE_SB_THREAD
 113 static void
 114 unlink_thread(struct thread *th)
 115 {
 116     if (th->prev)
 117         th->prev->next = th->next;
 118     else
 119         all_threads = th->next;
 120     if (th->next)
 121         th->next->prev = th->prev;
 122 }
 123 #endif
 124
 125 static int
 126 initial_thread_trampoline(struct thread *th)
 127 {
 128     lispobj function;
 129 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 130     lispobj *args = NULL;
 131 #endif
 132     function = th->no_tls_value_marker;
 133     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 134     if(arch_os_thread_init(th)==0) return 1;
 135     link_thread(th);
 136     th->os_thread=thread_self();
 137 #ifndef LISP_FEATURE_WIN32
 138     protect_control_stack_guard_page(1);
 139 #endif
 140
 141 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 142     return call_into_lisp_first_time(function,args,0);
 143 #else
 144     return funcall0(function);
 145 #endif
 146 }
 147
 148 #ifdef LISP_FEATURE_SB_THREAD
 149 #define THREAD_STATE_LOCK_SIZE \
 150     (sizeof(pthread_mutex_t))+(sizeof(pthread_cond_t))
 151 #else
 152 #define THREAD_STATE_LOCK_SIZE 0
 153 #endif
 154
 155 #define THREAD_STRUCT_SIZE (thread_control_stack_size + BINDING_STACK_SIZE + \
 156                             ALIEN_STACK_SIZE +                               \
 157                             THREAD_STATE_LOCK_SIZE +                         \
 158                             dynamic_values_bytes +                           \
 159                             32 * SIGSTKSZ +                                  \
 160                             THREAD_ALIGNMENT_BYTES)
 161
 162 #ifdef LISP_FEATURE_SB_THREAD
 163 /* THREAD POST MORTEM CLEANUP
 164  *
 165  * Memory allocated for the thread stacks cannot be reclaimed while
 166  * the thread is still alive, so we need a mechanism for post mortem
 167  * cleanups. FIXME: We actually have three, for historical reasons as
 168  * the saying goes. Do we really need three? Nikodemus guesses that
 169  * not anymore, now that we properly call pthread_attr_destroy before
 170  * freeing the stack. */
 171
 172 static struct thread_post_mortem *
 173 plan_thread_post_mortem(struct thread *corpse)
 174 {
 175     if (corpse) {
 176         struct thread_post_mortem *post_mortem = malloc(sizeof(struct thread_post_mortem));
 177         gc_assert(post_mortem);
 178         post_mortem->os_thread = corpse->os_thread;
 179         post_mortem->os_attr = corpse->os_attr;
 180         post_mortem->os_address = corpse->os_address;
 181 #ifdef DELAY_THREAD_POST_MORTEM
 182         post_mortem->next = NULL;
 183 #endif
 184         return post_mortem;
 185     } else {
 186         /* FIXME: When does this happen? */
 187         return NULL;
 188     }
 189 }
 190
 191 static void
 192 perform_thread_post_mortem(struct thread_post_mortem *post_mortem)
 193 {
 194 #ifdef CREATE_POST_MORTEM_THREAD
 195     pthread_detach(pthread_self());
 196 #endif
 197     if (post_mortem) {
 198         gc_assert(!pthread_join(post_mortem->os_thread, NULL));
 199         gc_assert(!pthread_attr_destroy(post_mortem->os_attr));
 200         free(post_mortem->os_attr);
 201         os_invalidate(post_mortem->os_address, THREAD_STRUCT_SIZE);
 202         free(post_mortem);
 203     }
 204 }
 205
 206 static void
 207 schedule_thread_post_mortem(struct thread *corpse)
 208 {
 209     struct thread_post_mortem *post_mortem = NULL;
 210     if (corpse) {
 211         post_mortem = plan_thread_post_mortem(corpse);
 212
 213 #ifdef DELAY_THREAD_POST_MORTEM
 214         pthread_mutex_lock(&thread_post_mortem_lock);
 215         /* First stick the new post mortem to the end of the queue. */
 216         if (pending_thread_post_mortem) {
 217             struct thread_post_mortem *next = pending_thread_post_mortem;
 218             while (next->next) {
 219                 next = next->next;
 220             }
 221             next->next = post_mortem;
 222         } else {
 223             pending_thread_post_mortem = post_mortem;
 224         }
 225         /* Then, if there are enough things in the queue, clean up one
 226          * from the head -- or increment the count, and null out the
 227          * post_mortem we have. */
 228         if (pending_thread_post_mortem_count > DELAY_THREAD_POST_MORTEM) {
 229             post_mortem = pending_thread_post_mortem;
 230             pending_thread_post_mortem = post_mortem->next;
 231         } else {
 232             pending_thread_post_mortem_count++;
 233             post_mortem = NULL;
 234         }
 235         pthread_mutex_unlock(&thread_post_mortem_lock);
 236         /* Finally run, the cleanup, if any. */
 237         perform_thread_post_mortem(post_mortem);
 238 #elif defined(CREATE_POST_MORTEM_THREAD)
 239         gc_assert(!pthread_create(&thread, NULL, perform_thread_post_mortem, post_mortem));
 240 #else
 241         post_mortem = (struct thread_post_mortem *)
 242             swap_lispobjs((lispobj *)(void *)&pending_thread_post_mortem,
 243                           (lispobj)post_mortem);
 244         perform_thread_post_mortem(post_mortem);
 245 #endif
 246     }
 247 }
 248
 249 /* this is the first thing that runs in the child (which is why the
 250  * silly calling convention).  Basically it calls the user's requested
 251  * lisp function after doing arch_os_thread_init and whatever other
 252  * bookkeeping needs to be done
 253  */
 254 int
 255 new_thread_trampoline(struct thread *th)
 256 {
 257     lispobj function;
 258     int result, lock_ret;
 259
 260     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 261     check_deferrables_blocked_or_lose();
 262     check_gc_signals_unblocked_or_lose();
 263     function = th->no_tls_value_marker;
 264     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 265     if(arch_os_thread_init(th)==0) {
 266         /* FIXME: handle error */
 267         lose("arch_os_thread_init failed\n");
 268     }
 269
 270     th->os_thread=thread_self();
 271     protect_control_stack_guard_page(1);
 272     /* Since GC can only know about this thread from the all_threads
 273      * list and we're just adding this thread to it, there is no
 274      * danger of deadlocking even with SIG_STOP_FOR_GC blocked (which
 275      * it is not). */
 276     lock_ret = pthread_mutex_lock(&all_threads_lock);
 277     gc_assert(lock_ret == 0);
 278     link_thread(th);
 279     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 280     gc_assert(lock_ret == 0);
 281
 282     result = funcall0(function);
 283
 284     /* Block GC */
 285     block_blockable_signals();
 286     set_thread_state(th, STATE_DEAD);
 287
 288     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 289      * thread, but since we are already dead it won't wait long. */
 290     lock_ret = pthread_mutex_lock(&all_threads_lock);
 291     gc_assert(lock_ret == 0);
 292
 293     gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
 294     unlink_thread(th);
 295     pthread_mutex_unlock(&all_threads_lock);
 296     gc_assert(lock_ret == 0);
 297
 298     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 299     pthread_mutex_destroy(th->state_lock);
 300     pthread_cond_destroy(th->state_cond);
 301
 302     os_invalidate((os_vm_address_t)th->interrupt_data,
 303                   (sizeof (struct interrupt_data)));
 304
 305 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 306     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 307     mach_port_move_member(mach_task_self(),
 308                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 309                           MACH_PORT_NULL);
 310     mach_port_deallocate(mach_task_self(),
 311                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 312     mach_port_destroy(mach_task_self(),
 313                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 314 #endif
 315
 316     schedule_thread_post_mortem(th);
 317     FSHOW((stderr,"/exiting thread %lu\n", thread_self()));
 318     return result;
 319 }
 320
 321 #endif /* LISP_FEATURE_SB_THREAD */
 322
 323 static void
 324 free_thread_struct(struct thread *th)
 325 {
 326     if (th->interrupt_data)
 327         os_invalidate((os_vm_address_t) th->interrupt_data,
 328                       (sizeof (struct interrupt_data)));
 329     os_invalidate((os_vm_address_t) th->os_address,
 330                   THREAD_STRUCT_SIZE);
 331 }
 332
 333 /* this is called from any other thread to create the new one, and
 334  * initialize all parts of it that can be initialized from another
 335  * thread
 336  */
 337
 338 static struct thread *
 339 create_thread_struct(lispobj initial_function) {
 340     union per_thread_data *per_thread;
 341     struct thread *th=0;        /*  subdue gcc */
 342     void *spaces=0;
 343     void *aligned_spaces=0;
 344 #ifdef LISP_FEATURE_SB_THREAD
 345     unsigned int i;
 346 #endif
 347
 348     /* May as well allocate all the spaces at once: it saves us from
 349      * having to decide what to do if only some of the allocations
 350      * succeed. SPACES must be appropriately aligned, since the GC
 351      * expects the control stack to start at a page boundary -- and
 352      * the OS may have even more rigorous requirements. We can't rely
 353      * on the alignment passed from os_validate, since that might
 354      * assume the current (e.g. 4k) pagesize, while we calculate with
 355      * the biggest (e.g. 64k) pagesize allowed by the ABI. */
 356     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 357     if(!spaces)
 358         return NULL;
 359     /* Aligning up is safe as THREAD_STRUCT_SIZE has
 360      * THREAD_ALIGNMENT_BYTES padding. */
 361     aligned_spaces = (void *)((((unsigned long)(char *)spaces)
 362                                + THREAD_ALIGNMENT_BYTES-1)
 363                               &~(unsigned long)(THREAD_ALIGNMENT_BYTES-1));
 364     per_thread=(union per_thread_data *)
 365         (aligned_spaces+
 366          thread_control_stack_size+
 367          BINDING_STACK_SIZE+
 368          ALIEN_STACK_SIZE +
 369          THREAD_STATE_LOCK_SIZE);
 370
 371 #ifdef LISP_FEATURE_SB_THREAD
 372     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 373         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 374     if (all_threads == 0) {
 375         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 376             SetSymbolValue
 377                 (FREE_TLS_INDEX,
 378                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 379                  make_fixnum(MAX_INTERRUPTS+
 380                              sizeof(struct thread)/sizeof(lispobj)),
 381                  0);
 382             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 383         }
 384 #define STATIC_TLS_INIT(sym,field) \
 385   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 386   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 387
 388         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 389         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 390         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 391         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 392         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 393 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 394         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 395 #endif
 396 #undef STATIC_TLS_INIT
 397     }
 398 #endif
 399
 400     th=&per_thread->thread;
 401     th->os_address = spaces;
 402     th->control_stack_start = aligned_spaces;
 403     th->binding_stack_start=
 404         (lispobj*)((void*)th->control_stack_start+thread_control_stack_size);
 405     th->control_stack_end = th->binding_stack_start;
 406     th->alien_stack_start=
 407         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 408     th->binding_stack_pointer=th->binding_stack_start;
 409     th->this=th;
 410     th->os_thread=0;
 411 #ifdef LISP_FEATURE_SB_THREAD
 412     th->os_attr=malloc(sizeof(pthread_attr_t));
 413     th->state_lock=(pthread_mutex_t *)((void *)th->alien_stack_start +
 414                                        ALIEN_STACK_SIZE);
 415     pthread_mutex_init(th->state_lock, NULL);
 416     th->state_cond=(pthread_cond_t *)((void *)th->state_lock +
 417                                       (sizeof(pthread_mutex_t)));
 418     pthread_cond_init(th->state_cond, NULL);
 419 #endif
 420     th->state=STATE_RUNNING;
 421 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 422     th->alien_stack_pointer=((void *)th->alien_stack_start
 423                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 424 #else
 425     th->alien_stack_pointer=((void *)th->alien_stack_start);
 426 #endif
 427 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 428     th->pseudo_atomic_bits=0;
 429 #endif
 430 #ifdef LISP_FEATURE_GENCGC
 431     gc_set_region_empty(&th->alloc_region);
 432 #endif
 433
 434 #ifndef LISP_FEATURE_SB_THREAD
 435     /* the tls-points-into-struct-thread trick is only good for threaded
 436      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 437      * appropriate values from struct thread here, and make sure that
 438      * we use the appropriate SymbolValue macros to access any of the
 439      * variable quantities from the C runtime.  It's not quite OAOOM,
 440      * it just feels like it */
 441     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 442     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 443     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 444 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 445     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 446     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 447     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 448 #else
 449     current_binding_stack_pointer=th->binding_stack_pointer;
 450     current_control_stack_pointer=th->control_stack_start;
 451 #endif
 452 #endif
 453     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 454     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 455     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 456     bind_variable(INTERRUPT_PENDING, NIL,th);
 457     bind_variable(INTERRUPTS_ENABLED,T,th);
 458     bind_variable(ALLOW_WITH_INTERRUPTS,T,th);
 459     bind_variable(GC_PENDING,NIL,th);
 460     bind_variable(ALLOC_SIGNAL,NIL,th);
 461 #ifdef LISP_FEATURE_SB_THREAD
 462     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 463 #endif
 464
 465     th->interrupt_data = (struct interrupt_data *)
 466         os_validate(0,(sizeof (struct interrupt_data)));
 467     if (!th->interrupt_data) {
 468         free_thread_struct(th);
 469         return 0;
 470     }
 471     th->interrupt_data->pending_handler = 0;
 472     th->interrupt_data->gc_blocked_deferrables = 0;
 473     th->no_tls_value_marker=initial_function;
 474
 475     th->stepping = NIL;
 476     return th;
 477 }
 478
 479 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 480 mach_port_t setup_mach_exception_handling_thread();
 481 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 482
 483 #endif
 484
 485 void create_initial_thread(lispobj initial_function) {
 486     struct thread *th=create_thread_struct(initial_function);
 487     if(th) {
 488 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 489         setup_mach_exception_handling_thread();
 490 #endif
 491         initial_thread_trampoline(th); /* no return */
 492     } else lose("can't create initial thread\n");
 493 }
 494
 495 #ifdef LISP_FEATURE_SB_THREAD
 496
 497 #ifndef __USE_XOPEN2K
 498 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 499                                   size_t __stacksize);
 500 #endif
 501
 502 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 503 {
 504     /* The new thread inherits the restrictive signal mask set here,
 505      * and enables signals again when it is set up properly. */
 506     sigset_t oldset;
 507     boolean r=1;
 508     int retcode = 0, initcode;
 509
 510     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 511
 512     /* Blocking deferrable signals is enough, no need to block
 513      * SIG_STOP_FOR_GC because the child process is not linked onto
 514      * all_threads until it's ready. */
 515     thread_sigmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 516
 517 #ifdef LOCK_CREATE_THREAD
 518     retcode = pthread_mutex_lock(&create_thread_lock);
 519     gc_assert(retcode == 0);
 520     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 521 #endif
 522
 523     if((initcode = pthread_attr_init(th->os_attr)) ||
 524        /* call_into_lisp_first_time switches the stack for the initial thread. For the
 525         * others, we use this. */
 526        (pthread_attr_setstack(th->os_attr,th->control_stack_start,thread_control_stack_size)) ||
 527        (retcode = pthread_create
 528         (kid_tid,th->os_attr,(void *(*)(void *))new_thread_trampoline,th))) {
 529         FSHOW_SIGNAL((stderr, "init = %d\n", initcode));
 530         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 531         if(retcode < 0) {
 532             perror("create_os_thread");
 533         }
 534         r=0;
 535     }
 536
 537 #ifdef LOCK_CREATE_THREAD
 538     retcode = pthread_mutex_unlock(&create_thread_lock);
 539     gc_assert(retcode == 0);
 540     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 541 #endif
 542     thread_sigmask(SIG_SETMASK,&oldset,0);
 543     return r;
 544 }
 545
 546 os_thread_t create_thread(lispobj initial_function) {
 547     struct thread *th, *thread = arch_os_get_current_thread();
 548     os_thread_t kid_tid = 0;
 549
 550     /* Must defend against async unwinds. */
 551     if (SymbolValue(INTERRUPTS_ENABLED, thread) != NIL)
 552         lose("create_thread is not safe when interrupts are enabled.\n");
 553
 554     /* Assuming that a fresh thread struct has no lisp objects in it,
 555      * linking it to all_threads can be left to the thread itself
 556      * without fear of gc lossage. initial_function violates this
 557      * assumption and must stay pinned until the child starts up. */
 558     th = create_thread_struct(initial_function);
 559     if (th && !create_os_thread(th,&kid_tid)) {
 560         free_thread_struct(th);
 561         kid_tid = 0;
 562     }
 563     return kid_tid;
 564 }
 565
 566 /* stopping the world is a two-stage process.  From this thread we signal
 567  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 568  * the usual pseudo-atomic checks (we don't want to stop a thread while
 569  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 570  */
 571
 572 /* To avoid deadlocks when gc stops the world all clients of each
 573  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 574  * holding the lock, but they must agree on which. */
 575 void gc_stop_the_world()
 576 {
 577     struct thread *p,*th=arch_os_get_current_thread();
 578     int status, lock_ret;
 579 #ifdef LOCK_CREATE_THREAD
 580     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 581      * on FreeBSD. */
 582     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock\n"));
 583     lock_ret = pthread_mutex_lock(&create_thread_lock);
 584     gc_assert(lock_ret == 0);
 585     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock\n"));
 586 #endif
 587     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock\n"));
 588     /* keep threads from starting while the world is stopped. */
 589     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 590     gc_assert(lock_ret == 0);
 591
 592     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock\n"));
 593     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 594     for(p=all_threads; p; p=p->next) {
 595         gc_assert(p->os_thread != 0);
 596         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: thread=%lu, state=%x\n",
 597                       p->os_thread, thread_state(p)));
 598         if((p!=th) && ((thread_state(p)==STATE_RUNNING))) {
 599             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending thread %lu\n",
 600                           p->os_thread));
 601             /* We already hold all_thread_lock, P can become DEAD but
 602              * cannot exit, ergo it's safe to use pthread_kill. */
 603             status=pthread_kill(p->os_thread,SIG_STOP_FOR_GC);
 604             if (status==ESRCH) {
 605                 /* This thread has exited. */
 606                 gc_assert(thread_state(p)==STATE_DEAD);
 607             } else if (status) {
 608                 lose("cannot send suspend thread=%lu: %d, %s\n",
 609                      p->os_thread,status,strerror(status));
 610             }
 611         }
 612     }
 613     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 614     for(p=all_threads;p;p=p->next) {
 615         if (p!=th) {
 616             FSHOW_SIGNAL
 617                 ((stderr,
 618                   "/gc_stop_the_world: waiting for thread=%lu: state=%x\n",
 619                   p->os_thread, thread_state(p)));
 620             wait_for_thread_state_change(p, STATE_RUNNING);
 621             if (p->state == STATE_RUNNING)
 622                 lose("/gc_stop_the_world: unexpected state");
 623         }
 624     }
 625     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 626 }
 627
 628 void gc_start_the_world()
 629 {
 630     struct thread *p,*th=arch_os_get_current_thread();
 631     int lock_ret;
 632     /* if a resumed thread creates a new thread before we're done with
 633      * this loop, the new thread will get consed on the front of
 634      * all_threads, but it won't have been stopped so won't need
 635      * restarting */
 636     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 637     for(p=all_threads;p;p=p->next) {
 638         gc_assert(p->os_thread!=0);
 639         if (p!=th) {
 640             lispobj state = thread_state(p);
 641             if (state != STATE_DEAD) {
 642                 if(state != STATE_SUSPENDED) {
 643                     lose("gc_start_the_world: wrong thread state is %d\n",
 644                          fixnum_value(state));
 645                 }
 646                 FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 647                               p->os_thread));
 648                 set_thread_state(p, STATE_RUNNING);
 649             }
 650         }
 651     }
 652
 653     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 654     gc_assert(lock_ret == 0);
 655 #ifdef LOCK_CREATE_THREAD
 656     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 657     gc_assert(lock_ret == 0);
 658 #endif
 659
 660     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 661 }
 662 #endif
 663
 664 int
 665 thread_yield()
 666 {
 667 #ifdef LISP_FEATURE_SB_THREAD
 668     return sched_yield();
 669 #else
 670     return 0;
 671 #endif
 672 }
 673
 674 /* If the thread id given does not belong to a running thread (it has
 675  * exited or never even existed) pthread_kill _may_ fail with ESRCH,
 676  * but it is also allowed to just segfault, see
 677  * <http://udrepper.livejournal.com/16844.html>.
 678  *
 679  * Relying on thread ids can easily backfire since ids are recycled
 680  * (NPTL recycles them extremely fast) so a signal can be sent to
 681  * another process if the one it was sent to exited.
 682  *
 683  * We send signals in two places: signal_interrupt_thread sends a
 684  * signal that's harmless if delivered to another thread, but
 685  * SIG_STOP_FOR_GC is fatal.
 686  *
 687  * For these reasons, we must make sure that the thread is still alive
 688  * when the pthread_kill is called and return if the thread is
 689  * exiting. */
 690 int
 691 kill_safely(os_thread_t os_thread, int signal)
 692 {
 693 #ifdef LISP_FEATURE_SB_THREAD
 694     sigset_t oldset;
 695     struct thread *thread;
 696     /* pthread_kill is not async signal safe and we don't want to be
 697      * interrupted while holding the lock. */
 698     thread_sigmask(SIG_BLOCK, &deferrable_sigset, &oldset);
 699     pthread_mutex_lock(&all_threads_lock);
 700     for (thread = all_threads; thread; thread = thread->next) {
 701         if (thread->os_thread == os_thread) {
 702             int status = pthread_kill(os_thread, signal);
 703             if (status)
 704                 lose("kill_safely: pthread_kill failed with %d\n", status);
 705             break;
 706         }
 707     }
 708     pthread_mutex_unlock(&all_threads_lock);
 709     thread_sigmask(SIG_SETMASK,&oldset,0);
 710     if (thread)
 711         return 0;
 712     else
 713         return -1;
 714 #else
 715     int status;
 716     if (os_thread != 0)
 717         lose("kill_safely: who do you want to kill? %d?\n", os_thread);
 718     status = raise(signal);
 719     if (status == 0) {
 720         return 0;
 721     } else {
 722         lose("cannot raise signal %d, %d %s\n",
 723              signal, status, strerror(errno));
 724     }
 725 #endif
 726 }